Multi-agent systems (MAS) combine multiple agents to solve complex tasks, and recent systems have grown to more than a thousand agents. At that scale, a central orchestrator that assigns sub-tasks becomes impractical, so decentralized designs let agents coordinate on their own.

AgentCue, described in a new arXiv paper, proposes using statistical summaries to guide this self-organization. Instead of receiving explicit sub-task assignments, agents would use these summaries to make local decisions and align their behavior with the overall system.

The paper's abstract introduces the concept but does not report specific experimental outcomes. The contribution, as presented, is the idea of statistical summaries as a lightweight coordination signal for large decentralized MAS.